MCP Server for Azure Log Analytics — gives Devin access to production logs. Implements the Model Context Protocol to expose Azure Log Analytics queries as tools that the Devin agent can invoke during incident investigation.
Four tools with basic schemas and descriptions, but significant gaps in parameter documentation, output schema definition, and error guidance. Tool names follow verb_noun convention and are generally clear. Descriptions exist but lack depth on when/why to use each tool. Parameters are defined with types but lack validation constraints, format specifications, and dependency documentation. No documented output schemas. Error handling is present but generic, returns dict with 'error' key without recovery guidance. Security: credentials properly injected via environment variables (good), but no audit logging or permission checks. Overall: serviceable but not production-grade.
Get recent exceptions from a specific service. Returns exception type, message, and stack trace.
Get error rate and request metrics for a specific service over the last hour.
Get recent log traces from a specific service, filtered by severity.
Execute a KQL query against Azure Log Analytics. Use this to investigate application logs, exceptions, and traces from the EventFlow services.
No documented output schema for any tool. LLMs cannot predict return structure or plan downstream calls. All four tools return 'list[dict]' but field names, types, and required keys are undocumented. Fixes dependency-driven composition (pattern:tool-chain).
query_logs parameter 'query' lacks validation guidance. No format specification, length limits, or examples of valid KQL syntax. LLM may hallucinate invalid Kusto queries. Should include pattern or reference to KQL documentation.
Error responses are generic: [{'error': 'message'}]. No recovery guidance. Examples: 'Log Analytics client not configured' and 'Query returned partial results' tell LLM nothing about next steps. Should provide recovery suggestions (e.g., 'Verify AZURE_LOG_ANALYTICS_WORKSPACE_ID is set' or 'Check query syntax with Azure KQL reference').
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 52 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 46 | 2024-11-05+ | v1 |
get_exceptions and get_traces construct KQL queries via string interpolation with user-provided 'service_name' parameter. Vulnerable to Kusto injection. Example: if service_name='foo" | take 1000000', the injected clause bypasses intended limits. Must sanitize service_name or use parameterized queries.
Result limits hardcoded in KQL (e.g., 'take 20', 'take 50', 'take 12') but not exposed as tool parameters. Agents cannot request fewer results to save tokens or more to explore further. Should expose 'limit' as a parameter with min/max bounds and document in description.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All four tools are read-only and idempotent, which should be declared in tool metadata. LLMs cannot infer from schema alone.
Parameter 'min_severity' in get_traces uses an integer enum (0=Verbose, 1=Info, 2=Warning, 3=Error, 4=Critical) but is not declared as an enum in schema. Description lists allowed values as prose, not a machine-readable enum. LLM may pass invalid integers like 5 or -1.
Timespan parameter in query_logs accepts 'ISO 8601 duration' but provides no validation or examples beyond 'PT1H'. LLM may pass 'PT30M', 'P1D', or invalid strings like '1 hour'. Should specify format constraints and provide 2-3 examples.
No audit logging. Calls to query_logs, get_exceptions, get_traces, get_metrics are logged only at INFO level with exception stacktraces. No record of which agent invoked what query, with what parameters, at what timestamp, and what the result was. Required for compliance and incident response.